CoverCraft Architecture: From Naive GenAI to Evidence-Grounded Application Intelligence
An architectural deep dive into technical honesty, zero-hallucination trust boundaries, deterministic application-layer scoring, real-time web grounding, and Model Context Protocol (MCP) orchestration.
The High-Level Architectural Vision
Standard generative AI tools treat job applications as creative writing exercises. When a candidate uploads their resume and a target job description, conventional tools pass them as raw prompt text to an LLM with instructions like "Write an impressive cover letter highlighting why I am a 95% match."
In reality, hiring managers, senior engineering reviewers, and automated screening systems evaluate applications through the lens of verifiable evidence. CoverCraft was designed from the ground up to replace creative text generation with an Evidence-Backed Application Intelligence Topology.
Strict Context Pipeline with Quarantined Ingestion
The Naive AI Wrapper Fallacy vs. Enterprise Evidence Architecture
When software engineers evaluate generative AI products, the primary question is: "Is this an ungrounded prompt wrapper around an OpenAI or Gemini endpoint, or is there an authentic software and verification architecture?"
| Architecture Dimension | Typical Naive AI Wrapper | CoverCraft Enterprise Architecture |
|---|---|---|
| Match Scoring | LLM asked to guess a percentage ("Match: 92%"). Non-deterministic, unrepeatable, hallucinates high scores. | Deterministic formula computed in application layer. LLM only categorizes evidence; math is calculated in code. |
| Candidate Claims | Model invents accomplishments ("Led team of 25", "Built scalable Kubernetes infra") not present in resume. | Strict <user_resume> quarantine. Every claim must have a verbatim quote. Unsupported claims are flagged as HIGH RISK. |
| Company Intel | Generic flattery ("I admire your innovative culture and passion for excellence"). Stale training data cutoff. | Real-time Tavily search engine integration. 4 structured cards with clickable [1], [2] citation pills to original web sources. |
| ATS Audit | Claims proprietary ATS score (e.g., "94/100 ATS Score"), which is deceptive since no ATS exposes private scoring algorithms. | Deterministic 6-point heuristic check (keyword density, format parseability, contact presence) with explicit technical disclaimer. |
| Tool Transparency | Black box opaque spinner. No visibility into system prompts, tool calls, or failure fallbacks. | Model Context Protocol (MCP) stream terminal. 6 registered tools logging execution events with expandable JSON schema payloads. |
Context Isolation, Sanitization & Prompt Security
CoverCraft decouples prompt logic into isolated policy files stored in src/prompts/. The architecture enforces three distinct context layers:
System Instructions
Passed strictly into the LLM's systemInstruction parameter. Developer rules: Anti-flattery, strict evidence requirement, JSON schema enforcement.
XML Boundary Quarantine
Raw candidate resumes and job descriptions are treated as untrusted inputs. They are wrapped in explicit boundary tags: <user_resume> and <job_description>.
Structured Output Enforcement
Enforces responseMimeType: "application/json". Raw unparseable text is rejected, preventing prompt injection bypasses.
🛡️ Document Ingestion & Sanitization Engine (/api/parse-resume)
Before parsing reaches the AI engine, uploaded files pass through a multi-stage security pipeline:
- Strict 5 MB Size Gate: Files exceeding 5,242,880 bytes are rejected immediately at the HTTP boundary.
- Multi-Format Parser: Native PDF extraction using
pdf-parseonUint8Arraybuffers; DOCX extraction usingmammoth; raw UTF-8 text decoding for TXT. - Null-Byte & Control Character Stripping: Eradicates null bytes (
- Prompt Injection Heuristic Scanner: Automatically scans for adversarial injection patterns (e.g., "ignore all previous instructions", "system prompt:", "<|im_start|>") and flags suspicious documents with a security warning.
- Character Quotas: Maximum 15,000 characters for resumes and 10,000 characters for job descriptions to prevent token exhaustion denial-of-service.
Technical Honesty: Deterministic Application-Layer Scoring
One of the most critical engineering decisions in CoverCraft is eliminating Black-Box LLM Score Hallucination. When an LLM is asked to output an overall fit percentage, it generates an impressionistic, non-reproducible number. CoverCraft removes scoring authority from the LLM entirely.
Categorical Classification to Mathematical Summation
Real-Time Web Grounding & The 4-Card Company Intel Dossier
Static LLMs suffer from knowledge cutoffs and cannot know a company's live hiring status, engineering priorities, or workplace reputation in 2026. CoverCraft integrates a deep real-time web retrieval and intelligence pipeline in src/app/api/research/route.js.
Deep multi-page crawl targeting company engineering, tech stack, and portal postings.
High-fidelity markdown scraping via r.jina.ai with zero JavaScript, cookie, or ad bloat.
Detects third-party staffing payrolls (C2H/agencies) with strict verbatim JD quotation proof.
Live Grounding to In-Text Interactive Citations
Whenever the model asserts a fact (e.g., "Google DeepMind developed the AMIE multi-agent clinical reasoning system [3]"), the citation token [3] is rendered as an interactive button. Clicking it opens a floating modal with the exact source title, category, relevance rating, snippet quote, and direct external URL.
Model Context Protocol (MCP) & Execution Transparency
CoverCraft implements the open Model Context Protocol (MCP) standard, exposing an extensible tool server in mcp-server/server.py and a streamable execution trace terminal in the frontend.
Extracts required competency vectors and separates hard technical prerequisites from nice-to-haves.
Executes real-time Tavily search queries for company developments and engineering initiatives.
Deduplicates URLs, removes SEO noise, and classifies domains into Official, Research, or News.
Synthesizes evidence-backed company dossier with explicit in-text citations.
Claim Ledger gatekeeper: Quarantines statements inside <user_resume> boundary.
Drafts 4-paragraph evidence-first cover letter using Gemini 3.5 with fallback to 3.8.
Adversarial Interview Defense Harness
Writing an impressive cover letter is only half the battle. If a candidate cannot defend the claims during a grueling technical or behavioral interview, the application fails. CoverCraft introduces an Adversarial Cross-Examination Engine (src/prompts/interview-defense.system.js).
The model assumes the role of an adversarial senior interviewer interrogating every claim in the generated cover letter. For each potential vulnerability:
Skills marked as transferable or partial where the candidate lacks direct production tenure.
Quantified metrics that an interviewer will probe for architecture details and individual contribution.
Well-established baseline qualifications verified by multiple resume projects.
Generative UI Architecture & Component Hydration
Legacy AI applications suffer from the "Markdown Wall" anti-pattern — returning unstructured text blocks that force users to copy-paste into an external editor. CoverCraft rejects this paradigm by implementing a full Generative UI Architecture. Instead of raw prose, the LLM emits structured telemetry payloads that the Next.js frontend dynamically compiles into interactive, stateful React components on the fly.
Footnote tags like [1] and [2] within the synthesized letter are not plain text. The client-side parser parses regex tokens and mounts clickable interactive badges that trigger an Evidence Drawer displaying domain attribution, publication dates, and verbatim source quotations.
Before synthesis begins, the pipeline renders live curation checkboxes across 4 intelligence cards. Users can inspect extracted Tavily search signals, veto any hallucinated or irrelevant company facts, and dictate exactly what evidence the letter synthesizer is permitted to cite.
Generated paragraphs hydrate into individual modular sections with real-time word counting, inline rich-text editing, and instant LaTeX-formatted PDF compilation via react-to-print. Candidates can customize tone sentence-by-sentence with zero layout shift.
The interview defense harness generates reactive flashcards equipped with risk tier badges (HIGH, MEDIUM, LOW), recruiter psychological intent analysis, and expandable "Anchored Resume Proof" snippets so candidates can rehearse live interview answers before walking into the room.
step: approval_gate → generating → results) → Interactive Generative UI components. Zero page refreshes.Codebase Audit & Technical Verification Directory
Every system, formula, boundary, and protocol described in this documentation maps directly to production code files in the repository:
Engineered by Ambuj Kumar Tripathi
Independent GenAI Engineer · Ex-British Telecom Automation · Gorakhpur, India
“The LLM is the least reliable part of your entire stack. That is why CoverCraft was engineered with deterministic ATS scoring, circuit breakers, and MCP tool verification instead of trusting raw generative models.”
Experience CoverCraft in Action
Upload your resume, analyze real job requirements, and inspect the live MCP trace.